add tsv110 pipeline scheduling
wuyuan (E)
wuyuan5@huawei.com
Thu Mar 7 14:25:00 GMT 2019
Hi ,James:
The modified patch has been uploaded for ten days. If you have time, I hope to get your comments earlier, thank you very much! Best Regards,
wuyuan
-----邮件原件-----
发件人: wuyuan (E)
发送时间: 2019年3月4日 21:46
收件人: 'James Greenhalgh' <james.greenhalgh@arm.com>
抄送: 'Kyrill Tkachov' <kyrylo.tkachov@foss.arm.com>; 'gcc-patches@gcc.gnu.org' <gcc-patches@gcc.gnu.org>; Zhangyichao (AB) <zhangyichao.zhang@huawei.com>; Zhanghaijian (A) <z.zhanghaijian@huawei.com>; 'nd@arm.com' <nd@arm.com>; wufeng (O) <wufeng9@huawei.com>; Yangfei (Felix) <felix.yang@huawei.com>
主题: re: add tsv110 pipeline scheduling
Hi ,James:
Have you seen the patch submitted last week? If the problem with the patch has been fixed, I hope to get into the trunk earlier. look forward to your reply. Thank you.
Best Regards,
wuyuan
-----邮件原件-----
发件人: wuyuan (E)
发送时间: 2019年2月23日 21:28
收件人: 'James Greenhalgh' <james.greenhalgh@arm.com>
抄送: Kyrill Tkachov <kyrylo.tkachov@foss.arm.com>; gcc-patches@gcc.gnu.org; Zhangyichao (AB) <zhangyichao.zhang@huawei.com>; Zhanghaijian (A) <z.zhanghaijian@huawei.com>; nd@arm.com; wufeng (O) <wufeng9@huawei.com>; Yangfei (Felix) <felix.yang@huawei.com> Re : add tsv110 pipeline scheduling
Hi ,James:
Sorry for not responding to your email in time because of Chinese New Year’s holiday and urgent work. The three questions you mentioned last email are due to my misunderstanding of pipeline.
the first question, These instructions will occupy both the tsv110_ls* and tsv110_fsu* Pipeline at the same time.
rewritten as follows:
(define_insn_reservation
"tsv110_neon_ld4_lane" 9
(and (eq_attr "tune" "tsv110")
(eq_attr "type" "neon_load4_all_lanes,neon_load4_all_lanes_q,\
neon_load4_one_lane,neon_load4_one_lane_q"))
"(tsv110_ls1 + tsv110_fsu1)|(tsv110_ls1 + tsv110_fsu2)|(tsv110_ls2 + tsv110_fsu1)|(tsv110_ls2 + tsv110_fsu2)")
the second question, These instructions will use tsv110_fsu1 Pipeline or tsv110_fsu2 Pipeline.
rewritten as follows:
(define_insn_reservation "tsv110_neon_abd_aba" 4
(and (eq_attr "tune" "tsv110")
(eq_attr "type" "neon_abd,neon_arith_acc"))
"tsv110_fsu1|tsv110_fsu2")
the third question, These instructions will use tsv110_fsu1 Pipeline or tsv110_fsu2 Pipeline.
rewritten as follows:
(define_insn_reservation "tsv110_neon_abd_aba_q" 4
(and (eq_attr "tune" "tsv110")
(eq_attr "type" "neon_arith_acc_q"))
"tsv110_fsu1|tsv110_fsu2")
In addition to the above changes, I asked hardware engineers and colleagues to review my patch and modify some of the errors. The detailed patches are as follows:
* config/aarch64/aarch64-cores.def (tsv1100): Change scheduling model.
* config/aarch64/aarch64.md : Add "tsv110.md"
* config/aarch64/tsv110.md: New file.
diff --git a/gcc/config/aarch64/aarch64-cores.def b/gcc/config/aarch64/aarch64-cores.def
index ed56e5e..82d91d6
--- a/gcc/config/aarch64/aarch64-cores.def
+++ b/gcc/config/aarch64/aarch64-cores.def
@@ -105,7 +105,7 @@ AARCH64_CORE("neoverse-n1", neoversen1, cortexa57, 8_2A, AARCH64_FL_FOR_ARCH8_ AARCH64_CORE("neoverse-e1", neoversee1, cortexa53, 8_2A, AARCH64_FL_FOR_ARCH8_2 | AARCH64_FL_F16 | AARCH64_FL_RCPC | AARCH64_FL_DOTPROD | AARCH64_FL_SSBS, cortexa53, 0x41, 0xd4a, -1)
/* HiSilicon ('H') cores. */
-AARCH64_CORE("tsv110", tsv110, cortexa57, 8_2A, AARCH64_FL_FOR_ARCH8_2 | AARCH64_FL_CRYPTO | AARCH64_FL_F16 | AARCH64_FL_AES | AARCH64_FL_SHA2, tsv110, 0x48, 0xd01, -1)
+AARCH64_CORE("tsv110", tsv110, tsv110, 8_2A, AARCH64_FL_FOR_ARCH8_2 | AARCH64_FL_CRYPTO | AARCH64_FL_F16 | AARCH64_FL_AES | AARCH64_FL_SHA2, tsv110, 0x48, 0xd01, -1)
/* ARMv8.4-A Architecture Processors. */
diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md index b7cd9fc..861f059 100644
--- a/gcc/config/aarch64/aarch64.md
+++ b/gcc/config/aarch64/aarch64.md
@@ -361,6 +361,7 @@
(include "thunderx.md")
(include "../arm/xgene1.md")
(include "thunderx2t99.md")
+(include "tsv110.md")
;; -------------------------------------------------------------------
;; Jumps and other miscellaneous insns
diff --git a/gcc/config/aarch64/tsv110.md b/gcc/config/aarch64/tsv110.md new file mode 100644 index 0000000..9d12839
--- /dev/null
+++ b/gcc/config/aarch64/tsv110.md
@@ -0,0 +1,708 @@
+;; tsv110 pipeline description
+;; Copyright (C) 2018 Free Software Foundation, Inc.
+;;
+;; This file is part of GCC.
+;;
+;; GCC is free software; you can redistribute it and/or modify it ;;
+under the terms of the GNU General Public License as published by ;;
+the Free Software Foundation; either version 3, or (at your option) ;;
+any later version.
+;;
+;; GCC is distributed in the hope that it will be useful, but ;;
+WITHOUT ANY WARRANTY; without even the implied warranty of ;;
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU ;;
+General Public License for more details.
+;;
+;; You should have received a copy of the GNU General Public License ;;
+along with GCC; see the file COPYING3. If not see ;;
+<http://www.gnu.org/licenses/>.
+
+(define_automaton "tsv110")
+
+(define_attr "tsv110_neon_type"
+ "neon_arith_acc, neon_arith_acc_q,
+ neon_arith_basic, neon_arith_complex,
+ neon_reduc_add_acc, neon_multiply, neon_multiply_q,
+ neon_multiply_long, neon_mla, neon_mla_q, neon_mla_long,
+ neon_sat_mla_long, neon_shift_acc, neon_shift_imm_basic,
+ neon_shift_imm_complex,
+ neon_shift_reg_basic, neon_shift_reg_basic_q, neon_shift_reg_complex,
+ neon_shift_reg_complex_q, neon_fp_negabs, neon_fp_arith,
+ neon_fp_arith_q, neon_fp_reductions_q, neon_fp_cvt_int,
+ neon_fp_cvt_int_q, neon_fp_cvt16, neon_fp_minmax, neon_fp_mul,
+ neon_fp_mul_q, neon_fp_mla, neon_fp_mla_q, neon_fp_recpe_rsqrte,
+ neon_fp_recpe_rsqrte_q, neon_fp_recps_rsqrts, neon_fp_recps_rsqrts_q,
+ neon_bitops, neon_bitops_q, neon_from_gp,
+ neon_from_gp_q, neon_move, neon_tbl3_tbl4, neon_zip_q, neon_to_gp,
+ neon_load_a, neon_load_b, neon_load_c, neon_load_d, neon_load_e,
+ neon_load_f, neon_store_a, neon_store_b, neon_store_complex,
+ unknown"
+ (cond [
+ (eq_attr "type" "neon_arith_acc, neon_reduc_add_acc,\
+ neon_reduc_add_acc_q")
+ (const_string "neon_arith_acc")
+ (eq_attr "type" "neon_arith_acc_q")
+ (const_string "neon_arith_acc_q")
+ (eq_attr "type" "neon_abs,neon_abs_q,neon_add, neon_add_q, neon_add_long,\
+ neon_add_widen, neon_neg, neon_neg_q,\
+ neon_reduc_add, neon_reduc_add_q,\
+ neon_reduc_add_long, neon_sub, neon_sub_q,\
+ neon_sub_long, neon_sub_widen, neon_logic,\
+ neon_logic_q, neon_tst, neon_tst_q,\
+ neon_compare, neon_compare_q,\
+ neon_compare_zero, neon_compare_zero_q,\
+ neon_minmax, neon_minmax_q, neon_reduc_minmax,\
+ neon_reduc_minmax_q")
+ (const_string "neon_arith_basic")
+ (eq_attr "type" "neon_add_halve_narrow_q,\
+ neon_add_halve, neon_add_halve_q,\
+ neon_sub_halve, neon_sub_halve_q, neon_qabs,\
+ neon_qabs_q, neon_qadd, neon_qadd_q, neon_qneg,\
+ neon_qneg_q, neon_qsub, neon_qsub_q,\
+ neon_sub_halve_narrow_q")
+ (const_string "neon_arith_complex")
+
+ (eq_attr "type" "neon_mul_b, neon_mul_h, neon_mul_s,\
+ neon_mul_h_scalar, neon_mul_s_scalar,\
+ neon_sat_mul_b, neon_sat_mul_h,\
+ neon_sat_mul_s, neon_sat_mul_h_scalar,\
+ neon_sat_mul_s_scalar,\
+ neon_mul_b_long, neon_mul_h_long,\
+ neon_mul_s_long,\
+ neon_mul_h_scalar_long, neon_mul_s_scalar_long,\
+ neon_sat_mul_b_long, neon_sat_mul_h_long,\
+ neon_sat_mul_s_long, neon_sat_mul_h_scalar_long,\
+ neon_sat_mul_s_scalar_long,\
+ neon_mla_b, neon_mla_h, neon_mla_s,\
+ neon_mla_h_scalar, neon_mla_s_scalar,\
+ neon_mla_b_long, neon_mla_h_long,\
+ neon_mla_s_long,\
+ neon_mla_h_scalar_long, neon_mla_s_scalar_long,\
+ neon_sat_mla_b_long, neon_sat_mla_h_long,\
+ neon_sat_mla_s_long, neon_sat_mla_h_scalar_long,\
+ neon_sat_mla_s_scalar_long")
+ (const_string "neon_multiply")
+ (eq_attr "type" "neon_mul_b_q, neon_mul_h_q, neon_mul_s_q,\
+ neon_mul_h_scalar_q, neon_mul_s_scalar_q,\
+ neon_sat_mul_b_q, neon_sat_mul_h_q,\
+ neon_sat_mul_s_q, neon_sat_mul_h_scalar_q,\
+ neon_sat_mul_s_scalar_q,\
+ neon_mla_b_q, neon_mla_h_q, neon_mla_s_q,\
+ neon_mla_h_scalar_q, neon_mla_s_scalar_q")
+ (const_string "neon_multiply_q")
+
+ (eq_attr "type" "neon_shift_acc, neon_shift_acc_q")
+ (const_string "neon_shift_acc")
+ (eq_attr "type" "neon_shift_imm, neon_shift_imm_q,\
+ neon_shift_imm_narrow_q, neon_shift_imm_long")
+ (const_string "neon_shift_imm_basic")
+ (eq_attr "type" "neon_sat_shift_imm, neon_sat_shift_imm_q,\
+ neon_sat_shift_imm_narrow_q")
+ (const_string "neon_shift_imm_complex")
+ (eq_attr "type" "neon_shift_reg")
+ (const_string "neon_shift_reg_basic")
+ (eq_attr "type" "neon_shift_reg_q")
+ (const_string "neon_shift_reg_basic_q")
+ (eq_attr "type" "neon_sat_shift_reg")
+ (const_string "neon_shift_reg_complex")
+ (eq_attr "type" "neon_sat_shift_reg_q")
+ (const_string "neon_shift_reg_complex_q")
+
+ (eq_attr "type" "neon_fp_neg_s, neon_fp_neg_s_q,\
+ neon_fp_abs_s, neon_fp_abs_s_q,\
+ neon_fp_neg_d, neon_fp_neg_d_q,\
+ neon_fp_abs_d, neon_fp_abs_d_q,\
+ neon_fp_minmax_s,neon_fp_minmax_d,\
+ neon_fp_reduc_minmax_s,neon_fp_reduc_minmax_d")
+ (const_string "neon_fp_negabs")
+ (eq_attr "type" "neon_fp_addsub_s, neon_fp_abd_s,\
+ neon_fp_reduc_add_s, neon_fp_compare_s,\
+ neon_fp_round_s,\
+ neon_fp_addsub_d, neon_fp_abd_d,\
+ neon_fp_reduc_add_d, neon_fp_compare_d,\
+ neon_fp_round_d")
+ (const_string "neon_fp_arith")
+ (eq_attr "type" "neon_fp_addsub_s_q, neon_fp_abd_s_q,\
+ neon_fp_reduc_add_s_q, neon_fp_compare_s_q,\
+ neon_fp_minmax_s_q, neon_fp_round_s_q,\
+ neon_fp_addsub_d_q, neon_fp_abd_d_q,\
+ neon_fp_reduc_add_d_q, neon_fp_compare_d_q,\
+ neon_fp_minmax_d_q, neon_fp_round_d_q")
+ (const_string "neon_fp_arith_q")
+ (eq_attr "type" "neon_fp_reduc_minmax_s_q,\
+ neon_fp_reduc_minmax_d_q,\
+ neon_fp_reduc_add_s_q, neon_fp_reduc_add_d_q")
+ (const_string "neon_fp_reductions_q")
+ (eq_attr "type" "neon_fp_to_int_s, neon_int_to_fp_s,\
+ neon_fp_to_int_d, neon_int_to_fp_d")
+ (const_string "neon_fp_cvt_int")
+ (eq_attr "type" "neon_fp_to_int_s_q, neon_int_to_fp_s_q,\
+ neon_fp_to_int_d_q, neon_int_to_fp_d_q")
+ (const_string "neon_fp_cvt_int_q")
+ (eq_attr "type" "neon_fp_cvt_narrow_s_q, neon_fp_cvt_widen_h")
+ (const_string "neon_fp_cvt16")
+ (eq_attr "type" "neon_fp_mul_s, neon_fp_mul_s_scalar,\
+ neon_fp_mul_d")
+ (const_string "neon_fp_mul")
+ (eq_attr "type" "neon_fp_mul_s_q, neon_fp_mul_s_scalar_q,\
+ neon_fp_mul_d_q, neon_fp_mul_d_scalar_q")
+ (const_string "neon_fp_mul_q")
+ (eq_attr "type" "neon_fp_mla_s, neon_fp_mla_s_scalar,\
+ neon_fp_mla_d")
+ (const_string "neon_fp_mla")
+ (eq_attr "type" "neon_fp_mla_s_q, neon_fp_mla_s_scalar_q,
+ neon_fp_mla_d_q, neon_fp_mla_d_scalar_q")
+ (const_string "neon_fp_mla_q")
+ (eq_attr "type" "neon_fp_recpe_s, neon_fp_rsqrte_s,\
+ neon_fp_recpx_s,\
+ neon_fp_recpe_d, neon_fp_rsqrte_d,\
+ neon_fp_recpx_d")
+ (const_string "neon_fp_recpe_rsqrte")
+ (eq_attr "type" "neon_fp_recpe_s_q, neon_fp_rsqrte_s_q,\
+ neon_fp_recpx_s_q,\
+ neon_fp_recpe_d_q, neon_fp_rsqrte_d_q,\
+ neon_fp_recpx_d_q")
+ (const_string "neon_fp_recpe_rsqrte_q")
+ (eq_attr "type" "neon_fp_recps_s, neon_fp_rsqrts_s,\
+ neon_fp_recps_d, neon_fp_rsqrts_d")
+ (const_string "neon_fp_recps_rsqrts")
+ (eq_attr "type" "neon_fp_recps_s_q, neon_fp_rsqrts_s_q,\
+ neon_fp_recps_d_q, neon_fp_rsqrts_d_q")
+ (const_string "neon_fp_recps_rsqrts_q")
+ (eq_attr "type" "neon_bsl, neon_cls, neon_cnt,\
+ neon_rev, neon_permute, neon_rbit,\
+ neon_tbl1, neon_tbl2, neon_zip,\
+ neon_dup, neon_dup_q, neon_ext, neon_ext_q,\
+ neon_move, neon_move_q, neon_move_narrow_q")
+ (const_string "neon_bitops")
+ (eq_attr "type" "neon_bsl_q, neon_cls_q, neon_cnt_q,\
+ neon_rev_q, neon_permute_q, neon_rbit_q")
+ (const_string "neon_bitops_q")
+ (eq_attr "type" "neon_from_gp,f_mcr,f_mcrr")
+ (const_string "neon_from_gp")
+ (eq_attr "type" "neon_from_gp_q")
+ (const_string "neon_from_gp_q")
+
+ (eq_attr "type" "f_loads, f_loadd,\
+ neon_load1_1reg, neon_load1_1reg_q,\
+ neon_load1_2reg, neon_load1_2reg_q")
+ (const_string "neon_load_a")
+ (eq_attr "type" "neon_load1_3reg, neon_load1_3reg_q,\
+ neon_load1_4reg, neon_load1_4reg_q")
+ (const_string "neon_load_b")
+ (eq_attr "type" "neon_load1_one_lane, neon_load1_one_lane_q,\
+ neon_load1_all_lanes, neon_load1_all_lanes_q,\
+ neon_load2_2reg, neon_load2_2reg_q,\
+ neon_load2_all_lanes, neon_load2_all_lanes_q")
+ (const_string "neon_load_c")
+ (eq_attr "type" "neon_load2_4reg, neon_load2_4reg_q,\
+ neon_load3_3reg, neon_load3_3reg_q,\
+ neon_load3_one_lane, neon_load3_one_lane_q,\
+ neon_load4_4reg, neon_load4_4reg_q")
+ (const_string "neon_load_d")
+ (eq_attr "type" "neon_load2_one_lane, neon_load2_one_lane_q,\
+ neon_load3_all_lanes, neon_load3_all_lanes_q,\
+ neon_load4_all_lanes, neon_load4_all_lanes_q")
+ (const_string "neon_load_e")
+ (eq_attr "type" "neon_load4_one_lane, neon_load4_one_lane_q")
+ (const_string "neon_load_f")
+
+ (eq_attr "type" "f_stores, f_stored,\
+ neon_store1_1reg")
+ (const_string "neon_store_a")
+ (eq_attr "type" "neon_store1_2reg, neon_store1_1reg_q")
+ (const_string "neon_store_b")
+ (eq_attr "type" "neon_store1_3reg, neon_store1_3reg_q,\
+ neon_store3_3reg, neon_store3_3reg_q,\
+ neon_store2_4reg, neon_store2_4reg_q,\
+ neon_store4_4reg, neon_store4_4reg_q,\
+ neon_store2_2reg, neon_store2_2reg_q,\
+ neon_store3_one_lane, neon_store3_one_lane_q,\
+ neon_store4_one_lane, neon_store4_one_lane_q,\
+ neon_store1_4reg, neon_store1_4reg_q,\
+ neon_store1_one_lane, neon_store1_one_lane_q,\
+ neon_store2_one_lane, neon_store2_one_lane_q")
+ (const_string "neon_store_complex")]
+ (const_string "unknown")))
+
+;; The tsv110 core is modelled as issues pipeline that has ;; the
+following functional units.
+;; 1. Three pipelines for integer operations: ALU1, ALU2, ALU3
+
+(define_cpu_unit "tsv110_alu1_issue" "tsv110") (define_reservation
+"tsv110_alu1" "tsv110_alu1_issue")
+
+(define_cpu_unit "tsv110_alu2_issue" "tsv110") (define_reservation
+"tsv110_alu2" "tsv110_alu2_issue")
+
+(define_cpu_unit "tsv110_alu3_issue" "tsv110") (define_reservation
+"tsv110_alu3" "tsv110_alu3_issue")
+
+;; 2. One pipeline for complex integer operations: MDU
+
+(define_cpu_unit "tsv110_mdu_issue" "tsv110") (define_reservation
+"tsv110_mdu" "tsv110_mdu_issue")
+
+;; 3. Two asymmetric pipelines for Asimd and FP operations: FSU1, FSU2
+(define_automaton "tsv110_fsu")
+
+(define_cpu_unit "tsv110_fsu1_issue"
+ "tsv110_fsu")
+(define_cpu_unit "tsv110_fsu2_issue"
+ "tsv110_fsu")
+
+(define_reservation "tsv110_fsu1" "tsv110_fsu1_issue")
+(define_reservation "tsv110_fsu2" "tsv110_fsu2_issue")
+
+;; 4. Two pipeline for branch operations but same with alu2 and alu3:
+BRU1, BRU2
+
+;; 5. Two pipelines for load and store operations: LS1, LS2.
+
+(define_cpu_unit "tsv110_ls1_issue" "tsv110") (define_cpu_unit
+"tsv110_ls2_issue" "tsv110") (define_reservation "tsv110_ls1"
+"tsv110_ls1_issue") (define_reservation "tsv110_ls2"
+"tsv110_ls2_issue")
+
+;; Block all issue queues.
+
+(define_reservation "tsv110_block" "tsv110_fsu1_issue + tsv110_fsu2_issue
+ + tsv110_mdu_issue + tsv110_alu1_issue
+ + tsv110_alu2_issue + tsv110_alu3_issue + tsv110_ls1_issue +
+tsv110_ls2_issue")
+
+;; Simple Execution Unit:
+;;
+;; Simple ALU without shift
+(define_insn_reservation "tsv110_alu" 1
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "alu_imm,logic_imm,\
+ alu_sreg,logic_reg,\
+ adc_imm,adc_reg,\
+ adr,bfm,clz,rbit,rev,\
+ shift_imm,shift_reg,\
+ mov_imm,mov_reg,\
+ mvn_imm,mvn_reg,\
+ mrs,multiple,no_insn"))
+ "tsv110_alu1|tsv110_alu2|tsv110_alu3")
+
+(define_insn_reservation "tsv110_alus" 1
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "alus_imm,logics_imm,\
+ alus_sreg,logics_reg,\
+ adcs_imm,adcs_reg"))
+ "tsv110_alu2|tsv110_alu3")
+
+;; ALU ops with shift
+(define_insn_reservation "tsv110_alu_shift" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "extend,\
+ alu_shift_imm,alu_shift_reg,\
+ crc,logic_shift_imm,logic_shift_reg,\
+ mov_shift,mvn_shift,\
+ mov_shift_reg,mvn_shift_reg"))
+ "tsv110_mdu")
+
+(define_insn_reservation "tsv110_alus_shift" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "alus_shift_imm,alus_shift_reg,\
+ logics_shift_imm,logics_shift_reg"))
+ "tsv110_alu2")
+
+;; Multiplies instructions
+(define_insn_reservation "tsv110_mult" 3
+ (and (eq_attr "tune" "tsv110")
+ (ior (eq_attr "mul32" "yes")
+ (eq_attr "widen_mul64" "yes")))
+ "tsv110_mdu")
+
+;; Integer divide
+(define_insn_reservation "tsv110_div" 10
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "udiv,sdiv"))
+ "tsv110_mdu")
+
+;; Block all issue pipes for a cycle
+(define_insn_reservation "tsv110_block" 1
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "block"))
+ "tsv110_block")
+
+;; Branch execution Unit
+;;
+;; Branches take two issue slot.
+;; No latency as there is no result
+(define_insn_reservation "tsv110_branch" 0
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "branch"))
+ "tsv110_alu2|tsv110_alu3")
+
+;; Load-store execution Unit
+;;
+;; Loads of up to two words.
+(define_insn_reservation "tsv110_load1" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "load_4,load_8"))
+ "tsv110_ls1|tsv110_ls2")
+
+;; Stores of up to two words.
+(define_insn_reservation "tsv110_store1" 0
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "store_4,store_8"))
+ "tsv110_ls1|tsv110_ls2")
+
+;; Advanced SIMD Unit - Integer Arithmetic Instructions.
+
+(define_insn_reservation "tsv110_neon_abd_aba" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_abd,neon_arith_acc"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation "tsv110_neon_abd_aba_q" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_arith_acc_q"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation "tsv110_neon_arith_basic" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_arith_basic"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation "tsv110_neon_arith_complex" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_arith_complex"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+;; Integer Multiply Instructions.
+;; D-form
+(define_insn_reservation "tsv110_neon_multiply" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_multiply"))
+ "tsv110_fsu1")
+
+(define_insn_reservation "tsv110_neon_multiply_dlong" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_mul_d_long"))
+ "tsv110_fsu1")
+
+;; Q-form
+(define_insn_reservation "tsv110_neon_multiply_q" 8
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_multiply_q"))
+ "tsv110_fsu1")
+
+;; Integer Shift Instructions.
+
+(define_insn_reservation
+ "tsv110_neon_shift_acc" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_shift_acc,\
+ neon_shift_imm_basic,neon_shift_imm_complex,neon_shift_reg_basic,\
+ neon_shift_reg_complex"))
+ "tsv110_fsu1")
+
+(define_insn_reservation
+ "tsv110_neon_shift_acc_q" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_shift_reg_basic_q,\
+ neon_shift_reg_complex_q"))
+ "tsv110_fsu1")
+
+;; Floating Point Instructions.
+
+(define_insn_reservation
+ "tsv110_neon_fp_negabs" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_fp_negabs"))
+ "(tsv110_fsu1|tsv110_fsu2)")
+
+(define_insn_reservation
+ "tsv110_neon_fp_arith" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_fp_arith"))
+ "(tsv110_fsu1|tsv110_fsu2)")
+
+(define_insn_reservation
+ "tsv110_neon_fp_arith_q" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_fp_arith_q"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_fp_minmax_q" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_fp_minmax_s_q,neon_fp_minmax_d_q"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_fp_reductions_q" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_fp_reductions_q"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_fp_cvt_int" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type"
+"neon_fp_cvt_int,neon_fp_cvt_int_q"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_fp_mul" 5
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_fp_mul"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_fp_mul_q" 5
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_fp_mul_q"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_fp_mla" 7
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_fp_mla,\
+ neon_fp_recps_rsqrts"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_fp_recpe_rsqrte" 3
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_fp_recpe_rsqrte"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_fp_mla_q" 7
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_fp_mla_q,\
+ neon_fp_recps_rsqrts_q"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_fp_recpe_rsqrte_q" 3
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_fp_recpe_rsqrte_q"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+;; Miscellaneous Instructions.
+
+(define_insn_reservation
+ "tsv110_neon_bitops" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_bitops"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_dup" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_from_gp,f_mcr"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_mov" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "f_mcrr"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_bitops_q" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_bitops_q"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_from_gp_q" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_from_gp_q"))
+ "(tsv110_alu1+tsv110_fsu1)|(tsv110_alu1+tsv110_fsu2)")
+
+(define_insn_reservation
+ "tsv110_neon_to_gp" 3
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_to_gp,neon_to_gp_q"))
+ "tsv110_fsu1")
+
+;; Load Instructions.
+
+(define_insn_reservation
+ "tsv110_neon_ld1_lane" 8
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_load1_one_lane,neon_load1_one_lane_q,\
+ neon_load1_all_lanes,neon_load1_all_lanes_q"))
+ "(tsv110_ls1 + tsv110_fsu1)|(tsv110_ls1 + tsv110_fsu2)|(tsv110_ls2 +
+tsv110_fsu1)|(tsv110_ls2 + tsv110_fsu2)")
+
+(define_insn_reservation
+ "tsv110_neon_ld1_reg1" 6
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type"
+"f_loads,f_loadd,neon_load1_1reg,neon_load1_1reg_q"))
+ "tsv110_ls1|tsv110_ls2")
+
+(define_insn_reservation
+ "tsv110_neon_ld1_reg2" 6
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_load1_2reg,neon_load1_2reg_q"))
+ "tsv110_ls1|tsv110_ls2")
+
+(define_insn_reservation
+ "tsv110_neon_ld1_reg3" 7
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_load1_3reg,neon_load1_3reg_q"))
+ "tsv110_ls1|tsv110_ls2")
+
+(define_insn_reservation
+ "tsv110_neon_ld1_reg4" 7
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_load1_4reg,neon_load1_4reg_q"))
+ "tsv110_ls1|tsv110_ls2")
+
+(define_insn_reservation
+ "tsv110_neon_ld2" 8
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_load1_2reg,neon_load1_2reg_q,\
+ neon_load2_2reg,neon_load2_2reg_q,neon_load2_all_lanes,\
+ neon_load2_all_lanes_q,neon_load2_one_lane,neon_load2_one_lane_q"))
+ "(tsv110_ls1 + tsv110_fsu1)|(tsv110_ls1 + tsv110_fsu2)|(tsv110_ls2 +
+tsv110_fsu1)|(tsv110_ls2 + tsv110_fsu2)")
+
+(define_insn_reservation
+ "tsv110_neon_ld3" 9
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_load3_3reg,neon_load3_3reg_q,\
+ neon_load3_one_lane,neon_load3_one_lane_q,\
+ neon_load3_all_lanes,neon_load3_all_lanes_q"))
+ "(tsv110_ls1 + tsv110_fsu1)|(tsv110_ls1 + tsv110_fsu2)|(tsv110_ls2 +
+tsv110_fsu1)|(tsv110_ls2 + tsv110_fsu2)")
+
+(define_insn_reservation
+ "tsv110_neon_ld4_lane" 9
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_load4_all_lanes,neon_load4_all_lanes_q,\
+ neon_load4_one_lane,neon_load4_one_lane_q"))
+ "(tsv110_ls1 + tsv110_fsu1)|(tsv110_ls1 + tsv110_fsu2)|(tsv110_ls2 +
+tsv110_fsu1)|(tsv110_ls2 + tsv110_fsu2)")
+
+(define_insn_reservation
+ "tsv110_neon_ld4_reg" 11
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "neon_load4_all_lanes,neon_load4_all_lanes_q,\
+ neon_load4_one_lane,neon_load4_one_lane_q"))
+ "(tsv110_ls1 + tsv110_fsu1)|(tsv110_ls1 + tsv110_fsu2)|(tsv110_ls2 +
+tsv110_fsu1)|(tsv110_ls2 + tsv110_fsu2)")
+
+;; Store Instructions.
+
+(define_insn_reservation
+ "tsv110_neon_store_a" 0
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_store_a"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation
+ "tsv110_neon_store_b" 0
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_store_b"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+;; These block issue for a number of cycles proportional to the number
+;; of 64-bit chunks they will store, we don't attempt to model that ;;
+precisely, treat them as blocking execution for two cycles when ;;
+issued.
+(define_insn_reservation
+ "tsv110_neon_store_complex" 0
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "tsv110_neon_type" "neon_store_complex"))
+ "tsv110_block*2")
+
+;; Floating-Point Operations.
+
+(define_insn_reservation "tsv110_fp_const" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "fconsts,fconstd,fmov"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation "tsv110_fp_add_sub" 5
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "fadds,faddd,fmuls,fmuld"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation "tsv110_fp_mac" 7
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "fmacs,ffmas,fmacd,ffmad"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation "tsv110_fp_cvt" 3
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "f_cvt"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation "tsv110_fp_cvtf2i" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "f_cvtf2i"))
+ "tsv110_fsu1")
+
+(define_insn_reservation "tsv110_fp_cvti2f" 5
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "f_cvti2f"))
+ "(tsv110_alu1+tsv110_fsu1)|(tsv110_alu1+tsv110_fsu2)")
+
+(define_insn_reservation "tsv110_fp_cmp" 4
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "fcmps,fcmpd"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation "tsv110_fp_arith" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "ffariths,ffarithd"))
+ "tsv110_fsu1|tsv110_fsu2")
+
+(define_insn_reservation "tsv110_fp_divs" 12
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "fdivs,neon_fp_div_s,fdivd,neon_fp_div_d,\
+ neon_fp_div_s_q,neon_fp_div_d_q"))
+ "tsv110_fsu1")
+
+(define_insn_reservation "tsv110_fp_sqrts" 24
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "fsqrts,neon_fp_sqrt_s,fsqrtd,neon_fp_sqrt_d,\
+ neon_fp_sqrt_s_q,neon_fp_sqrt_d_q"))
+ "tsv110_fsu2")
+
+(define_insn_reservation "tsv110_crypto_aes" 3
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "crypto_aese,crypto_aesmc"))
+ "tsv110_fsu1")
+
+(define_insn_reservation "tsv110_crypto_sha1_fast" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "crypto_sha1_fast,crypto_sha1_xor"))
+ "(tsv110_fsu1|tsv110_fsu2)")
+
+(define_insn_reservation "tsv110_crypto_sha256_fast" 2
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "crypto_sha1_fast"))
+ "tsv110_fsu1")
+
+(define_insn_reservation "tsv110_crypto_complex" 5
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "crypto_sha1_slow,crypto_sha256_slow"))
+ "tsv110_fsu1")
+
+;; We lie with calls. They take up all issue slots, but are otherwise
+;; not harmful.
+(define_insn_reservation "tsv110_call" 1
+ (and (eq_attr "tune" "tsv110")
+ (eq_attr "type" "call"))
+ "tsv110_alu1_issue+tsv110_alu2_issue+tsv110_alu3_issue+tsv110_fsu1_issue+tsv110_fsu2_issue\
+ +tsv110_mdu_issue+tsv110_ls1_issue+tsv110_ls2_issue"
+)
+
+;; Simple execution unit bypasses
+(define_bypass 1 "tsv110_alu"
+ "tsv110_alu,tsv110_alu_shift") (define_bypass 2
+"tsv110_alu_shift"
+ "tsv110_alu,tsv110_alu_shift")
+
+;; An MLA or a MUL can feed a dependent MLA.
+(define_bypass 3 "tsv110_neon_*mla*,tsv110_neon_*mul*"
+ "tsv110_neon_*mla*")
+
+;; We don't need to care about control hazards, either the branch is ;;
+predicted in which case we pay no penalty, or the branch is ;;
+mispredicted in which case instruction scheduling will be unlikely to
+;; help.
+(define_bypass 1 "tsv110_*"
+ "tsv110_call,tsv110_branch")
-----邮件原件-----
发件人: James Greenhalgh [mailto:james.greenhalgh@arm.com]
发送时间: 2019年1月18日 7:47
收件人: wuyuan (E) <wuyuan5@huawei.com>
抄送: Kyrill Tkachov <kyrylo.tkachov@foss.arm.com>; gcc-patches@gcc.gnu.org; Zhangyichao (AB) <zhangyichao.zhang@huawei.com>; Zhanghaijian (A) <z.zhanghaijian@huawei.com>; nd@arm.com
主题: Re: add tsv110 pipeline scheduling
On Mon, Jan 14, 2019 at 08:02:45AM -0600, wuyuan (E) wrote:
> Hi Kyrill:
> The gcc 7.3.0 does not discard the store1 and load1 command; I did not expect the community's latest gcc changes so large .
> now I downloaded the latest GCC code, put the patch into GCC source code, the compiler can pass, thank you very much for your work!
> Best Regards,
> wuyuan
Please check your modeling of Advanced SIMD operations.
> +(define_insn_reservation
> + "tsv110_neon_ld4_lane" 9
> + (and (eq_attr "tune" "tsv110")
> + (eq_attr "type" "neon_load4_all_lanes,neon_load4_all_lanes_q,\
> + neon_load4_one_lane,neon_load4_one_lane_q"))
> + "((tsv110_ls1*8)|(tsv110_ls2*8)|(tsv110_fsu1*8)|(tsv110_fsu2*8))")
> +
This model says you will reserve
LS1 for 8 cycles,
OR LS2 for 8 cycles,
OR FSU1 for 8 cycles,
OR FSU2 for 8 cycles.
> +(define_insn_reservation "tsv110_neon_abd_aba" 4
> + (and (eq_attr "tune" "tsv110")
> + (eq_attr "type" "neon_abd,neon_arith_acc"))
> + "tsv110_fsu1,tsv110_fsu2")
This model says you will reserve
FSU1 for 1 cycle,
THEN FSU2 for 1 cycle.
> +(define_insn_reservation "tsv110_neon_abd_aba_q" 4
> + (and (eq_attr "tune" "tsv110")
> + (eq_attr "type" "neon_arith_acc_q"))
> + "(tsv110_fsu1,tsv110_fsu2)+(tsv110_fsu1,tsv110_fsu2)")
> +
This model says you will reserve:
FSU1 for 1 cycle,
THEN FSU2 for 1 cycle
AND
FSU1 for 1 cycle,
THEN FSU2 for 1 cycle
Which would be a redundant AND.
Is that how you intend to model these operations?
Remember,
If you are looking to model a 'THEN' relationship you can use the ',' operator, If you are looking to model an 'AND' relationship you can use the '+' operator, If you are looking to model an 'OR' relationship you can use the '|' operator.
Taking Cortex-A57 as an example:
> (define_insn_reservation
> "cortex_a57_neon_load_d" 11
> (and (eq_attr "tune" "cortexa57")
> (eq_attr "cortex_a57_neon_type" "neon_load_d"))
> "ca57_cx1_issue+ca57_cx2_issue,
> ca57_ls_issue+ca57_ls_issue,ca57_ldr*2")
This model says you will reserve:
CX1_ISSUE AND CX2_ISSUE,
THEN LS_ISSUE AND LS_ISSUE,
THEN LDR for 2 cycles.
Please let me know if you plan to update the model. If I have misunderstood your intentions, please accept my apologies.
Best Regards,
James Greenhalgh
>
>
> * config/aarch64/aarch64-cores.def (tsv1100): Change scheduling model.
> * config/aarch64/aarch64.md : Add "tsv110.md"
> * config/aarch64/tsv110.md: New file.
More information about the Gcc-patches
mailing list